English · 简体中文

MemoryHub

Manage every AI coding session as git-backed memory — save it purified, load it back, curate it on a map.

MemoryHub (mh) is a small Python CLI and a local web map for Claude Code, pi and Codex. A session ends and mh save keeps it; the next session runs mh load and the project's memory is back; mh ui is where you manage all of it.

• Every session kept, none of the noise — purified to the User/Agent dialog by rule, no model call, stored as a file in a checkpoint, committed to a git repo you own
• Loaded when it matters — the next session starts from the checkpoint you are at, its parents and its links, newest first within a token budget
• Curated, not hoarded — skip a session, drop or rewrite an exchange, move a session, write a summary instead, from the map or the terminal
• Shaped like the project — stages from a template, parallel takes at a stage, sub-checkpoints for a smaller scope, links between the ones that belong together

uv tool install memoryhub-mh && mh skill install

Linux or macOS · git ≥ 2.32 · Python ≥ 3.12 · pipx install memoryhub-mh and pip install memoryhub-mh work too. See the changelog for what is new.

See it in action

Real captures of the map on a small café-site project, not mockups.

Click a node	Open a checkpoint	Save the running session
Everything you can do to it, with a word on each.	Untick a session and every load leaves it out.	Dialog as it is, or a summary the agent writes.

A saved session is the dialog and nothing else. Edit or delete an exchange; the hub keeps the history.

The session running right now, thinking and tool calls included, with the save box on top.

New checkpoint: the template's next stage, another take at this one, a sub-checkpoint, or any name.

Quick start

1. Install

uv tool install memoryhub-mh   # or: pipx install memoryhub-mh · pip install memoryhub-mh
mh skill install               # teaches Claude Code the /mh workflow

The package is memoryhub-mh, the command is mh. uv fetches Python 3.12 when the machine has none; pipx and pip need it installed.

2. Give the project a hub

cd my-site
mh init --template frontend   # .memoryhub/ with the stages of a frontend project
mh checkpoint                 # the first stage: requirement-analysis

3. Work, save, load

# … work with Claude Code; at the end the agent runs:
mh save                       # this session, purified, into the checkpoint
# next time:
mh load                       # the memory, back in context
mh ui                         # the map

mh hook install makes Claude Code run the load at session start and the save at the end.

Manage sessions

Why MemoryHub

• Purified, mechanically — tool calls, thinking, harness wrappers and the unanswered last question are stripped by rule. What is stored reads like the conversation, and costs nothing to produce.
• A git repo, not a database — .memoryhub/ is plain markdown in a normal repository: diff it, push it, revert it, read it without mh.
• Independent by default — checkpoints load alone unless you link them; a sub-checkpoint loads inside its parents; a take is a parallel path, not a copy.
• The map tells the truth — purple is exactly what the next mh load packs; a link the load does not reach is grey; a terminal change shows up within a poll.
• Local only — loopback, a per-run token, no cloud. Nothing leaves the machine unless you push the hub. The one model call there is, the summary, uses the CLI you already run.

How it works

Commands

Hands-free

mh hook install            # this project: load at session start, save at the end
mh hook install --user     # every project
mh hook install --remove   # undo

SessionStart injects mh load; SessionEnd and PreCompact run mh save.

Reference and scope

• CONTRIBUTING.md — the invariants a change must keep · CHANGELOG.md — what changed when
• scripts/showcase.py rebuilds the screenshots above from a throwaway project
• Out of scope on purpose: a hosted service, typing into the running session, choosing the template from the map

License

MIT — free to use, modify and distribute.

Contributing

Issues and pull requests are welcome. Start with the contribution guide; every change keeps the suite hermetic and both READMEs in step.